VSLAM method based on object detection in dynamic environments.

Area of Science:

• Computer Vision
• Robotics
• Augmented Reality

Background:

• Current visual simultaneous localization and mapping (VSLAM) systems struggle with dynamic environments, leading to pose estimation errors and tracking loss.
• Augmented Reality (AR) registration demands robust SLAM systems capable of handling complex and changing scenes.

Purpose of the Study:

• To develop a real-time tracking and mapping method that overcomes the limitations of traditional VSLAM in dynamic environments.
• To improve the accuracy and robustness of AR registration through enhanced SLAM capabilities.

Main Methods:

• The proposed method enhances the ORB-SLAM2 framework by integrating Gaussian Mixture Models (GMM) for background modeling and dynamic foreground segmentation.
• It utilizes the YOLOv3 object detector to identify dynamic targets within segmented regions.
• An improved Kalman filter tracks dynamic objects, and feature points are filtered to remove dynamic elements before map building.

Main Results:

• The method demonstrates improved robustness in dynamic datasets compared to standard VSLAM algorithms.
• Real-time augmented reality registration experiments showed stable and accurate placement of virtual objects.
• Pose estimation errors and camera tracking loss were significantly reduced in challenging dynamic scenes.

Conclusions:

• The combined GMM and YOLOv3 approach provides a more reliable solution for real-time tracking and mapping in dynamic environments.
• This enhanced SLAM method significantly improves the stability and accuracy of augmented reality registration.